PREreview of Spike inference from mouse spinal cord calcium imaging data

This study explores how well spike inference algorithms can decode neuronal activity from calcium imaging data in mouse spinal cord neurons, making a significant contribution in the field. By combining simultaneous electrophysiological recordings and calcium imaging, the researchers compared the performance of two algorithms: the supervised deep-learning-based CASCADE and the unsupervised OASIS, providing valuable insights into the strengths and limitations of the approaches. While CASCADE, initially trained on cortical neuron datasets, showed reasonable performance with spinal cord neurons, retraining it with ground-truth data specific to spinal glutamatergic and GABAergic neurons significantly enhanced its accuracy, particularly for high-frequency spike events. The study underscores the critical role of region-specific data in optimizing spike inference algorithms and presents retrained models that improve the interpretation of calcium imaging across diverse experimental contexts.

Major issues

• Methods: The number of animals used in each experimental group is not clearly stated. Please clarify. 

• The biological variability in calcium signal amplitudes and kinetics across different neurons is acknowledged but not comprehensively explored. Significant differences between GABAergic and glutamatergic neurons in terms of calcium transient amplitudes, decay times, and spike dynamics are evident in the figures but the manuscript would benefit from providing detailed biological interpretation. We suggest providing a deeper discussion on how these variations in calcium dynamics might influence algorithm performance and highlight whether differences in buffering capacity, channel types, or neuronal morphology are potential contributors.

• While the study demonstrates improved performance of the CASCADE algorithm after retraining with spinal cord-specific data, it does not sufficiently address whether these findings can generalize to other non-cortical regions or neuron types. The results focus exclusively on glutamatergic and GABAergic neurons in the spinal cord, but there is no discussion of how the retrained models might perform in other regions or under varying experimental conditions. Including a discussion on the potential limitations of generalizing these findings to other regions or neuronal subtypes and suggesting future experiments to evaluate broader applicability would enhance the manuscript.

• Although the authors explore noise levels and their impact on algorithm performance, their analysis remains limited in depth and does not provide a clear guideline for handling high-noise datasets. Providing a more in-depth analysis of noise impact and proposing practical approaches to mitigate its effects on spike inference would enhance the reach of the manuscript. 

• Discussion: The broader significance of the study’s findings for the field of neuroscience is not sufficiently addressed, particularly given the persistent error rates in spike inference. While the retraining process improved algorithm performance for spinal cord calcium imaging, the overall error rates remain high, potentially limiting the applicability of these algorithms for high-precision studies.The manuscript would benefit from discussing how the improved algorithms could be used in real-world scenarios, such as understanding spinal cord physiology or advancing clinical research, while acknowledging the challenges posed by residual errors. Proposing potential next steps to enhance algorithm reliability and reduce errors would further improve significance and citability of the paper.

Minor issues

• Contextual references. To provide broader context and highlight the relevance of spinal cord-specific findings, references to studies on calcium imaging in other non-cortical regions should be added. 

• Methodological Details. Please provide additional details on the retraining process for CASCADE, such as specific parameters and challenges encountered, to improve transparency and reproducibility.

• Clarity of figures. Figure legends, particularly for Figures 3 and 5, could be improved by including more detailed explanations of the experimental setup, parameters, and key observations.

• Figure 1. Smoothed spike rate (SR) traces are lacking a scale bar. We suggest the authors to consider converting this figure into a Supplementary Figure.

• Figure 7. We suggest using a more homogeneous criterion for the reference dF/F0 percentages (330 and 660%).

• Statistical reports. The article could benefit from stating the statistical tests performed within the legend of the figures, even though they are mentioned in the results. 

• Algorithm parameters. Please provide a clear table or appendix summarizing the key parameters used in training and testing both CASCADE and OASIS to facilitate reproducibility.

Competing interests

The authors declare that they have no competing interests.